The Herculaneum Excavation Hypermedia
Silvia Rossi
, Vincenzo Scognamiglio
and Ernesto Burattini
Dip. di Scienze Fisiche, University of Naples ”Federico II”, Via Cinthia, I80126 Naples, Italy
Centro di Ateneo per i Servizi Informativi, University of Naples ”Federico II”
Web Personalization, User Modeling, Metamodeling.
The aim of this work is the design and the implementation of a prototype for supporting a user through the
navigation in a web-site. We exploit the potentiality of having a representation of an hypermedia based on the
concept of an abstract semantic description of the knowledge tree, in order to use some adaptivity processes,
typical of many hypermedia systems, without the limitation of having a fixed set of resources. Our prototype
was implemented for the archaeological Herculaneum Excavations, but the novelty of this work is an attempt to
combine methodologies coming from Adaptive Hypermedia, Recommendation Systems and Semantic Web in
order to have a context-independent approach for Hypermedia on the Web. What we propose is an architecture
with modules that can be reused and adapted in many different contexts. In order to have a context independent
behavior the system was implemented with a structural adaptivity itself. This means that changing the tree the
system adapts itself to the modifications and dynamically regenerates the code of its components.
Nowadays several sites offer personalized portals that
can be customized by the user and may even adapt
to his interest automatically. In this framework there
are many applications that include personal guides for
navigation or ambient devices, portals, e-commerce
Web-sites (e.g. Amazon, Expedia), or recommender
sites (e.g MovieLens).
Usually, in web applications, one can identify two
different classes of adaptivity. In the first class, called
Adaptive Hypermedia, the adaptivity process is gen-
erated ad hoc for a specific system. Such systems
have precise information on a fixed structure of re-
sources, expressed as the content of each resource
and the relationships between them (Goldstein, 1979).
These solutions are hardly generalizable for re-use
in different applications. A reason why this adap-
tive functionality is not re-usable is related to the so–
called open corpus problem in adaptive hypermedia
(Brusilovsky, 2004). Adaptive Hypermedia Systems
are realized by experts that decide how to connect all
the information. That makes difficult to add new ma-
terial and to connect it.
In the second class, that uses Web Mining tech-
niques, the adaptivity is created in order to operate
in a network of Web resources. The are many ap-
proaches for creating adaptive web sites (Perkowitz
and Etzioni, 2001), but the only adaptive systems
that work successfully are the so called Web Recom-
mendation Systems (Brusilovsky, 2004). Those sys-
tems may exploit adaptation starting from the word-
level content of a resource (Content-based filtering)
or recording information of other users with similar
interests (collaborative filtering (Rashid et al., 2002)).
However the adaptivity process in those systems con-
sists in a list of recommended links that may be or-
dered according to some ratings. Moreover, an im-
portant aspect of Semantic Web (Berners-Lee et al.,
2001) is to obtain more and more weight for building
effective personalization in web applications. A mod-
ern system, in order to show flexibility and a context
independent behavior, needs to have a semantic de-
scription of resources. Such description may not be
bound to each single resource but can be expressed in
a more abstract way.
The aim of this work is the design and the imple-
mentation of a prototype for supporting a user through
Rossi S., Scognamiglio V. and Burattini E. (2007).
In Proceedings of the Third International Conference on Web Information Systems and Technologies - Web Interfaces and Applications, pages 270-275
DOI: 10.5220/0001283402700275
the navigation in a web-site. The adaptivity process is
obtained without asking any information to the user
that has to interact with the system as a common
web-site. Our prototype was implemented for the ar-
chaeological Herculaneum Excavations
(Sec.2), but
the novelty of this work is the attempt to combine
methodologies coming from Adaptive Hypermedia,
Recommendation Systems and Semantic Web in or-
der to have a context-independent approach that man-
ages Hypermedia on the web. WyW06 (Whatever)
Web You Want was born as general purpose adap-
tive system based on user modelling that can be used
as an information system on a huge set of data, but
in his evolution it becomes different from a classic
adaptive system, getting much powerful functionali-
ties typical of recommendation systems. In Sec.2.1
we will introduce our way to represent the knowledge
of a hypermedia, based on semantic description of the
knowledge tree, while the representation of resources
and of user models will be described in Sec.2.2 and
Sec.2.3. In Sec.2.4 and Sec.3 we will introduce the
adaptivity process in WyW06 and we will describe
how the system works.
The Herculaneum Excavations Hypermedia comes
from a stand-alone system by our research group
(Burattini et al., 1999). Herculaneum, as Pompeii,
was destroyed in A.D. 79 the 24th August by heavy
shower of lava from the Vesuvius volcano. The
knowledge base contains many information includ-
ing historical data and mythological tales, excavation
reports, description of recovered buildings and ob-
jects, painting and so on. This information comes
under the form of texts, videos, images, contempo-
rary and old maps. The knowledge was appropri-
ately linked following the experts’ suggestions
and the results of this analysis have been summarized
in a tree where single chunks of knowledge are iden-
tified (see Sec.2.1). A root node represents one of the
main points of view from which the knowledge on
Herculaneum Excavations may be explored and sub-
nodes show the ways in which such knowledge may
expanded. This first implementation has shown that,
even when a great amount of information is managed,
the system can be used only from a narrow class of
domain expert users since it is unable to adapt its in-
formation to different kinds of users. The current sys-
tem, realized as a web site, is intended for users whose
profiles can be different, such as archaeologists, his-
torians, onlookers or just tourists.
2.1 A Semantic Representation of the
Knowledge Tree
One of the main reasons for the not re-usability of
many adaptive system is the so-called open corpus
problem. First of all, we have to remark that each
adaptive hypermedia, as any hypermedia system, has
to make some assumptions about documents and their
relationships in a document space. Many of the adap-
tive applications work on a fixed set of documents that
are defined and linked at design time. The process of
adding a resource in the system always requires the
agency of domain experts and makes more complex
the maintenance of the system.
What we propose is an architecture that is context
independent and has modules that can be reused and
adapted. Our system has no a priori knowledge of the
web contents, but is able to manage a high level rep-
resentation defined on this set. The only requirement
we impose is that the resources are expressed using
some semantic annotation (see Sec.2.2) and that ex-
ists a high level representation in terms of a knowl-
edge tree. Web pages or, more generally, any hy-
permedia content, are generated dynamically starting
from those two representations. When the system has
to organize the data in order to provide a chunk of
information, or to create a web page, it collects the
appropriate contents from the database.
In order to clarify this point we have to make a
distinction among the concept of resource class and
the concept of resource instance. A content element
in the database represents a resource instance. For
example, in the case of the Herculaneum Excavation,
a resource instance may be the ”excavation report” of
a particular building or the ”description of a fresco”.
Another example of a resource instance in a different
context, such as travel recommendation systems, may
be the information of a specific hotel. Such instance
gets the semantic description of its class.
A resource class specifies a set of resource in-
stances that have similar semantic properties (see Fig.
1(d)). All the resource classes are semantically inter-
connected in order to represent the knowledge tree.
For example, in the case of the Herculaneum Excava-
tion, the abstract class building is connected to the
class graffiti in the sense that a building may have
graffiti. Another example concerning travel recom-
mendation system may be the abstract class town con-
nected to the class hotels, to the class tours, and so on.
At this point we claim that this type of abstract
knowledge representation can be easily used in many
Figure 1: (a) Virtual relationships generated by a knowledge
tree that is inserted in the database;
(b) To an instantiated tree corresponds a subset of resources;
(c) To a single resource corresponds a node in one or more
instantiated tree;
(d) To more resources in the database may correspond the
same node in the knowledge tree.
web applications. For example any recommendation
system may be enhanced by the use of this seman-
tic structure. Those systems, in fact, already have, in
an implicit way, a semantic representation of how the
global knowledge has to be organized and proposed.
The knowledge tree defines in an explicit way also the
resources and so the contents that we want to person-
alize and adapt to the user (see Fig.1). For instance
some of the nodes that are the leaves of the tree rep-
resent a more specific and technical content that can
be hidden for a user that is not an expert in the field.
Starting from the description of the knowledge tree
the system can make a search on the database and can
make different instances of the tree using the avail-
able resources. In this way we can create virtual re-
lationships among the resources in the database that
are not fixed a priori. Any modification of the knowl-
edge tree will automatically imply a modification of
the virtual links among the resources and the process
of adding a new resource does not need the interven-
tion of domain expert to connect it to the others. In
fact when a user requests a web page the system will
dynamically instantiate the correspondent virtual tree
and recognize all the new resources. The resource
tree for the Herculaneum Excavations is represented
in Fig.2 where each node represents a resource class.
2.2 Representation of the Resources
With the help of a group of architects, historians and
archaeologists we detected the main properties that
may characterize any information content of our hy-
permedia. Those features are related to the types of
informative content that a resource may have and to
the interests that a resource may arise. For example, a
feature is the historical content - i.e. how much a re-
source class contains contents coming from historical
data. A resource class is a vector
r = (w
, ..., w
) in
the space R
, where n is the number of characteris-
Figure 2: The knowledge tree for the Herculaneum Excava-
tions Hypermedia.
tics or features (w) and the orthogonal base is repre-
sented by the vectors that has a single feature with
value 100% and the other features with value 0%.
The vector model of a resource class represents how
much every specific characteristic is present in the
resource (see Fig.3). Differently from the common
usage in data mining the vector, that represents a re-
source, does not contain the frequency of occurrence
of some specific words within the text, but represents
how much a specific text can be classified according
to some typologies of texts. This can be interpreted
also as an analysis of the semantics of the keywords
within the text.
Once the vectors for the resource classes are de-
fined, any resource instance will inherit the model
from its correspondent class; this means that all the
instances of a class have the same semantic proper-
ties. A resource is an atomic and content-closed in-
formative kernel. Each hypermedia page is viewed as
specific subset of resources (see Sec.3).
<?xml version="1.0"?> <rdf:RDF xmlns:rdf=
<rdf:Description rdf:about=
<models:context>Archeological Cultural Heritage</context>
<models:nameOfResource>History of Excavations - archive
</rdf:Description> </rdf:RDF>
Figure 3: An example of a vector class in RDF.
In this work we are not dealing with the prob-
lem of resources’ classification. Resource models are
created by domain-knowledge experts and the addi-
WEBIST 2007 - International Conference on Web Information Systems and Technologies
tion of a node in knowledge tree implies the necessity
of defining a new resource model. Resource models
are static through the user interaction but can be eas-
ily changed by administrator that only has to update
the resource model. The system automatically re-
loads each redefinition when new user session starts.
Also the creation of new hypermedia pages, as con-
tainers of resources, is obtained dynamically by the
system just using the relationships described in the
2.3 User Model
In many classic adaptive hypermedia systems the
classification of the user is made through the use of
stereotypes (Kobsa, 1993), or by clustering a user
with a group of people that has shown the same be-
havior during past interactions. In the first case,
stereotypes must be incorporated in the system in the
designing phase and, during the execution, the system
will have to recognize the correct stereotype (Rich,
1999). In the second case the algorithms for cluster-
ing and the adaptivity process i.e. the recommen-
dation process – are exploit just recording the already
visited web pages or the rankings made by previous
users and they need a large amount of data in order to
be effective.
Taking the advantage of the semantic descriptions
of resources and of the knowledge tree, we decided to
implement a user model that takes into account only
the current behavior of the user starting from the spec-
ification of the behaviors of ”ideal” users. What is
different from the classical use of stereotypes is that
we do not try to classify a user in a specific class but
we start from the assumption that a real use may ex-
hibit a behavior that is a combination of ideal classes.
An ideal users’ class is the representation of a set of
users, like a prototype, whose behaviors and interests
are formally defined. It is fundamental that each ideal
class of users does not have any behavior that is in
common with another ideal class, i.e. the vectors rep-
resenting ideal classes are an orthogonal base.
A user model is a vector
u in the space of ideal
user classes. When a user session start, the user model
vector components are set to zero. During the user in-
teraction, the choices made by the user lead the sys-
tem to change the values of the user model. The user
model vectors in WyW06 are represented as couples
attribute–value and such values represent the percent-
age of similarity of the user model to an ideal user
class (see Fig.4). This is to say that the user model
of the user i is a linear combination of ideal user’s
models (
), such as
= α
, β
, ..., γ
, where
Greek letters represent percentages.
The content-independency of the system, obtained
for resource classes, is kept also in the management
of the user model. In WyW06 this vector represents a
user model in an explicit way, starting from the defin-
ition of ideal classes. The modification or the adding
of a new ideal class implies an automatic adaptation
of the system to the new configuration, without the
need of implementing new code since it is generated
by the system itself.
<?xml version="1.0"?> <rdf:RDF
<rdf:Description rdf:about=
<models:context>Archeological Cultural Heritage</context>
</rdf:Description> </rdf:RDF>
Figure 4: An example of a user model vector in RDF.
2.4 Extension and Association Rules:
Up to now, we described the user model and the re-
source model as two distinct concepts, without men-
tioning how to go from one representation to the other.
The user model changes according to the choices that
the user made browsing the resources, and the adap-
tivity process on resources depends on the current
user model. In Sec.2.2 we proposed the user model
as a composition of orthogonal ideal user classes. As-
sociated with each ideal class there is the specification
of the ”interests” of an ideal user. This specification
is made by using association rules. In the other way
the updating of the user model is made though the ap-
plication of extension rules e.g. rules that specify
how to modify the user model after a specific event.
Usually these rules are specified using statistical
methodologies on real data. In our case an associa-
tion rule is represented by the definition of a resource
class (
) considered optimal for an ideal user (
Such class is the one that has the typological content
more similar to the user interests. Moreover, the re-
sources, selected to be representative for the ideal user
models, are linearly independent, in order to respect
the fundamental assumption that the ideal users rep-
resent an orthogonal base (see Sec.2.3). To verify the
consistency of the user model representation, the sys-
tem, whenever there is a change in the ideal users’
specifications, calculates scalar products among the
resources associate to the ideal users and verifies that
they are orthogonal.
When a user requires a new hypermedia page, the
system takes the user model, computed during the in-
teraction up to that moment, and calculates the ideal
resource for the current user (
). Then it compares
a distance (di f f ) respect to each real resource (
) in
the selected page and, using a threshold for the ob-
tained distance, makes decision about which content
of the resource it has to show or hide. This process
is summarized in the following algorithm. Moreover
we want to remark that an ideal resource may not cor-
respond to a real resource.
Algorithm 1 DIFFERENCE(
user model
, resource class
= (w
, ..., w
= (α
, β
, ..., γ
1 for j = 1, j = n, j + +
2 r
i j
= max(αr
1 j
, βr
2 j
, ..., γr
m j
i j
is the j component of the ideal resource
r for the user i. r
i j
is the maximum value of all the j components of the resources
associated to the ideal classes of users
and multiplied by the
same factors as in the user model
3 di f f = 0
4 for all
secondaryNodesO f (
For all the secondary children of the node
for j = 1, j = n, j + +
For all the features of a node
5 if (r
i j
> r
k j
The user is not satisfied
6 di f f = di f f + 1
7 distance = 100 (
di f f
Compute the percentage of fields that are not satisfied
8 if (distance <= threshold)
9 show(
10 else
11 hide(
The metrics we use to compute the distance starts
from the assumption that a feature of a real resource
”satisfies” the user interests if his value is equal or
greater than the value of the same feature in the ideal
resource. This is to say that the real resource is ap-
pealing for the user at least for this specific feature.
For example if one has an interest of 20% for the his-
torical property, every real resource with the histori-
cal feature greater than 20% will satisfy his interest.
For each real resource, within a web page, the sys-
tem computes how many features do not satisfy the
user interests. If this number is low the content will
be shown otherwise it will be hidden. In future work
we are planning an evaluation of this metrics. It is im-
portant to highlight that the system lets the user free
to show the content of resources considered not in-
teresting for him and to hide the content of resource
automatically showed.
Finally, as we said, associated with each resource
there is an extension rule that specifies how to changes
the user model when the resource is activated. This
rule refers to the nodes of knowledge–tree but not
Figure 5: A snapshot of the prototype of the Herculaneum
Excavation web hypermedia.
to the resource instances. The adaptivity process in
WyW06 is achieved just by analyzing the user’s in-
teractions with the system environment without any
explicit questions about preferences.
WyW06 introduced a new and potentially powerful
layer of adaptation: the structural-adaptation. This
structural adaptation refers to the ability of a system
to create or regenerate some of his modules, dynami-
cally and autonomously, when some specific parame-
ters are changed.
To translate the tree in a hypermedia with his rel-
ative hyper–linked structure, WyW06 makes a fun-
damental and dynamical distinction between main
nodes and secondary nodes. How to make this choice
is relatively important, anyway, the current choice is
to evaluate the number of child of a specific node. If
it is smaller than a defined threshold then this node is
classified as secondary, a main node otherwise. The
addition of new instances of a resource does not re-
quire neither a complex intervention of system man-
agers the system automatically integrates the new
resource – nor the intervention of knowledge-domain
expert – the semantic of any new element refers to his
specific class model. The addition of a new node in
the tree is not absolutely an issue for the system that
regenerates parts of code of his modules and the hy-
pertext structures without other interventions.
At the beginning of a session each main resource
will appear as a web page with his main resource con-
tent and the list of resources that are children of the
main. If a main resource is a child of another, it will
appear as a link to another web page in the left box
(see Fig.5). If a secondary resource is child of the
main resource, it will appear as a expandable text
With the term expandable text we mean a subsection of
WEBIST 2007 - International Conference on Web Information Systems and Technologies
(see Fig.5) that user can expand in every moment.
When user expands or hides a secondary resource, its
associated extension rule is applied to the user model.
When the user model is sufficiently defined for the
system criteria and a user accesses to a main resource,
the system automatically expands the content of sec-
ondary resources that are considered interesting for
him. The system also provides to the user the pos-
sibility of saving resources as personal notes. This
action also implies the application of extension rules
to the user model.
For obtain the multilevel adaptation we imple-
mented the system using different language on server-
side and client-side. We used JavaScript combined
with Cascade Style Sheets (CSS) to realize the dy-
namic update of the fields of the user-model, resulting
by the application of extension rules, and the user’s
choice about viewing/hiding resources on client-side.
For the server-side, we used the PHP language to
obtain the dynamic generation by the server of the
contents–selection for the web pages. The PHP code
has the important role of applying the association
rules in order to obtain the user-model transformation.
The structural adaptation is made representing web
site structure, and relative resources, by knowledge-
tree. The real resources, the rules and the knowledge-
tree are stored in a database set up in MySQL envi-
ronment managed dynamically by PHP code.
WyW06 was born as general purpose adaptive sys-
tem based on user model. Like classic adaptive sys-
tem WyW06 has the aim of supporting individual
user in finding, selecting and managing content of-
fered by web sites. In his evolution WyW06 becomes
quickly more than a classic adaptive system, getting
much functionality typical of recommendation sys-
tems, melting the advantages of both those typolo-
gies of systems. WyW06 is based on the point of
view that each single web page is like a collection of
resources. A resource is an atomic informative ker-
nel semantically linked to other resources by a spe-
cific tree of relationships. The use of a tree to spec-
ify relationships between resources is a peculiar char-
acterization of adaptive hypermedia systems (Burke,
2002). In the adaptive hypermedia a common prob-
lem is the adding of a node-resource in the system.
As we said in section 2.1 to overcome this obstacle,
WyW06 introduced a new and potentially powerful
layer of adaptation: the structural-adaptation. The
the web page that can be hidden or shown by a click on a
structural adaptation refers to the ability of system
to create or regenerate some of his modules, dynam-
ically and autonomously, when some specific para-
meters are changed. Respect to classical definitions
in adaptive systems WyW06 is a user model based
system. The adaptive behavior of the system aims
to satisfy automatically the desires of the user but,
it is important to highlight that, lets the user free to
access to the resources that are not near to the self-
selection threshold. The use of such techniques leads
the system easily to become a recommendation sys-
tem. In fact the system evaluates the secondary re-
sources that are of more interest respect to the ac-
tual state of the user model and propose them to the
user. Summarizing we can say that WyW06 has a be-
havior comparable to an adaptive hypermedia, some
peculiarity of content-based recommendation system,
some characteristics of inferential reasoning by using
rules schemas, some fundamental notion for Seman-
tic Web and something innovative respect all, that is
the structural adaptive layer.
Berners-Lee, T., Hendler, J., and Lassila, O. (2001). The
semantic web. Scientific Am., pages 34–43.
Brusilovsky, P. (2004). Adaptive navigation support: From
adaptive hypermedia to the adaptive web and beyond.
Psychnology, 2(1):7–23.
Burattini, E., Gaudino, F., and Serino, L. (1999). Hyper-
media knowledge acquisition and a bdi agent for nav-
igation assistance. a case study: Herculaneum exca-
vations. In Europ. Conf. on Cognitive Science, pages
Burke, R. (2002). Hybrid recommender systems: Survey
and experiments. User Modeling and User-Adapted
Interaction, 12(4):331–370.
Goldstein, I. P. (1979). The genetic graph: A representation
for the evolution of procedural knowledge. Interna-
tional Journal of Man-Machine Studies, 11:51–77.
Kobsa, A. (1993). User modeling: Recent work, prospects
and hazards. In Adaptive User Interfaces: Principles
and Practice, pages 111–128. North-Holland, Ams-
Perkowitz, M. and Etzioni, O. (2001). Adaptive web sites:
Concept and case study. Artificial Intelligence, 118(1–
Rashid, A. M., Albert, I., Cosley, D., Lam, S. K., McNee,
S. M., Konstan, J. A., and Riedl, J. (2002). Getting
to know you: learning new user preferences in rec-
ommender systems. In IUI ’02: Proc. of the 7th int.
conf. on Intelligent user interfaces, pages 127–134,
New York, NY, USA. ACM Press.
Rich, E. (1999). Users are individuals: individualizing user
models. International Journal of Human Computer
Studies, 51(2):323–338.